AGORA
Adversarial Generation Of Real-time Animatable 3D Gaussian Head Avatars

Ramazan Fazylov1, Sergey Zagoruyko2, Aleksandr Parkin3, Stamatis Lefkimmiatis3, Ivan Laptev1
1MBZUAI, Abu Dhabi, UAE, 2Polynome AI, Dubai, UAE, 3MTS AI, Moscow, Russia
🎉 Accepted to the ECCV 2026 Interactive Social Avatars (ISA) Workshop 🎉
arXiv Code

Abstract

The generation of high-fidelity, animatable 3D human avatars remains a core challenge in computer graphics and vision, with applications in VR, telepresence, and entertainment. Existing approaches based on implicit representations like NeRFs suffer from slow rendering and dynamic inconsistencies, while 3D Gaussian Splatting (3DGS) methods are typically limited to static head generation, lacking dynamic control.

We bridge this gap by introducing AGORA, a novel framework that extends 3DGS within a generative adversarial network to produce animatable avatars. Our key contribution is a lightweight, FLAME-conditioned deformation branch that predicts per-Gaussian residuals, enabling identity-preserving, fine-grained expression control while allowing real-time inference. Identity is further preserved through spatial shape conditioning of the identity branch, and expression fidelity is enforced via a dual-discriminator training scheme leveraging synthetic renderings of the parametric mesh.

AGORA generates avatars that are not only visually realistic but also precisely controllable. Quantitatively, we outperform state-of-the-art NeRF-based methods on expression accuracy while rendering at 250 FPS on a single GPU and, notably, at ~9 FPS under CPU-only inference – to our knowledge the first demonstration of CPU-only animatable 3DGS avatar synthesis. This work represents a significant step toward practical, high-performance digital humans.

Method Overview

AGORA builds upon the UV-based GGHEAD framework, extending it into a fully animatable 3D GAN. The generator produces canonical 3DGS attributes — position, scale, rotation, color, opacity — in UV space from a latent code and FLAME shape parameters. A separate lightweight deformation branch takes low-resolution feature maps from the main generator and predicts per-Gaussian attribute residuals conditioned on FLAME expression ψ and jaw pose θ, which are composed with the canonical attributes to produce the final animated avatar. Gaussian positions are obtained by 3D lifting: interpolating base positions from the articulated FLAME mesh and adding predicted offsets, anchoring the 3DGS to the parametric mesh geometry.

Spatial Shape Conditioning

A key challenge in animatable generation is injecting shape priors (e.g., craniofacial proportions) without collapsing identity diversity. Naively injecting the FLAME shape code β into the mapping network causes the intermediate latent to be dominated by β, suppressing z-driven variation. Instead, we derive a UV-aligned map of the shape-isolated deformation field — the difference between a posed FLAME mesh with a given shape and a canonical reference — apply per-sample variance normalization, and concatenate it with the generator's block features. This injects shape biases spatially, where they matter geometrically, while preserving the stochasticity of the latent code.

Enforcing Expression Consistency

Training with only an image-based discriminator is insufficient for precise expression control. Following Next3D, we condition the discriminator on the target expression by concatenating the rendered image with a synthetic FLAME mesh rendering. Crucially, we replace UV-coordinate vertex coloring with a displacement-based signal: vertices are colored by their expression-isolated displacement from the neutral pose, giving the discriminator a fine-grained geometric cue to penalize expression deviations. This dual-discrimination scheme significantly improves mouth articulation and high-intensity expression fidelity.

Generated Avatars

Generated avatars (seeds 0-32) reenacted by the driving video on the left.

Single Image Avatars

Avatars generated from single images, driven by the video on the left.

More In-the-Wild Avatar Results

Comparison of avatar reenactment methods. Left: Ours, Center: Driving video, Right: Next3D.

Acknowledgements

This work builds upon GGHEAD, whose UV-based 3D Gaussian generation framework served as the foundation for our approach. We are grateful for their excellent codebase and insights.

We also acknowledge EG3D for its pioneering work in 3D-aware generative models, and Next3D, which inspired key aspects of our methodology. We thank the authors of concurrent work GAIA for providing visual results for comparison.

We are also grateful to the developers of gsplat for their efficient and easy-to-use Gaussian splatting library, which we use for rasterization.

BibTeX

@article{fazylov2025agora,
    author = {Fazylov, Ramazan and Zagoruyko, Sergey and Parkin, Aleksandr and Lefkimmiatis, Stamatis and Laptev, Ivan},
    title = {{AGORA: Adversarial Generation Of Real-time Animatable 3D Gaussian Head Avatars}},
    journal = {arXiv preprint arXiv:2512.06438v4},
    year = {2025},
    url = {https://arxiv.org/abs/2512.06438v4}
}